Deep learning techniques for satellite image classification

Abstract

Because of its wide range of uses in computer vision applications, including image retrieval, remote sensing, object recognition, scene analysis, and surveillance, image classification has attracted a lot of attention. Assigning appropriate class labels to images according to their contents is the primary objective of image classification. In the domain of remote sensing, image classification and analysis play crucial roles in both military and civil applications. Conventional methods for scene analysis and remote sensing depended on low-level representations of features, such as those of color and texture. However, recent advancements have shifted towards the use of convolutional neural networks (CNNs), which have shown promising results in remote sensing and scene classification tasks. In light of effectiveness of deep learning (DL) models, this research aims to develop four DL models by fine tuning already existing DL models-CNNs, residual neural network (ResNet), visual geometry group (VGG-19), network mobile net V2 based model and classifies satellite images of RSI-CB256 data set in to four classes namely cloudy, desert, green_area and water. For the RSI-CB256 dataset, appropriate network structures are explored in this research to get good performance. The CNN, ResNet and VGG-19 base models achieved an accuracy of 90.48, 92.68 and 91.18 respectively. While the mobile net V2 based model outperformed the other three models with 96.83% accuracy.